Towards Generating Novel Games Using Conceptual Blending

Experimental AI in Games: Papers from the AIIDE 2015 Workshop

Jeremy Gow and Joseph Corneli Computational Creativity Group Department of Computing Goldsmiths, University of London, UK

Abstract blending” (Schorlemmer et al. 2014)1. The project takes Goguen’s formalisation of conceptual blending as a starting We sketch the process of creating a novel video game by blending two video games specified in the Video Game point (Goguen 1999). Early implementations draw on the Description Language (VGDL), following the COIN- related notion of amalgams (Ontanõn´ and Plaza 2010). The VENT computational model of conceptual blending. core ideas in the model are: We highlight the choices that need to be made in this Conceptual spaces may be mathematical theories, ontolo- process, and discuss the prospects for a computational game designer based on blending. gies, or, in our case, video game descriptions. Signature morphisms between conceptual spaces are Introduction mappings from the symbols of a source conceptual space into the symbols of target conceptual space. In the This paper outlines the process of creating a novel video diagram below, φ1, φ2, ψ1, ψ2, and ι are all signature game by blending two video game descriptions given in the morphisms. Video Game Description Language (VGDL). Several issues need to be considered: which games to blend? Which el- Generic spaces are a particular kind of conceptual space ements of the game will be blended and which removed? that possess maximum commonality with a number of Do we need to adjust the resulting game in order to make it given input spaces. In the diagram below, φ1 and φ2 rep- playable? Taking two games from the GVG-AI Competition resent the inclusion of the elements of the generic space (Perez et al. 2015), we provide a worked example based on G in both X and Y . a recent computational, domain-independant model of con- Blends are unique conceptual spaces that preserve structure ceptual blending (Bou et al. 2015). from the input spaces and a generic space G, via a specific Our account provides a high-level sketch of how a com- arrangement of signature morphisms; namely, such that putational system could blend game designs, and does not the diagram is commutative, and that ‘axioms’ in the input propose a concrete architecture. Instead, our intent is ex- spaces X,Y are sent to ‘theorems’ in the unique colimit plore how a game blending system could work. We high- B, known as the blend: light the specific challenges presented by various steps in the blending process, and describe how certain steps provide B ψ ψ the opportunity to add or close off new potential mechanics 1 2 and meanings. The implementation and evaluation of this X ι Y approach will be the focus of future work. φ φ 1 G 2 Background Within the COINVENT project the model has been Recent research (Treanor et al. 2012; Cook and Colton 2014) primarily applied to examples from mathematics (Bou has examined computational generation of games from pro- et al. 2015) and music (Cambouropoulos, Kaliakatsos- vided concepts. The current exploration of blending game Papakostas, and Tsougras 2014). The constituent processes descriptions is motivated by our observation that game de- have been automated to differing degrees, and for some sim- signers often build new games by combining ideas from ex- ple examples complete automation is feasible. One of the isting games. key outstanding issues is the evaluation of blends, both from Conceptual blending is the notion that new concepts can the point of view of judging the final outcome, and from an be generated by combining elements of existing concepts. ‘online’ point of view that can help constrain the potential COINVENT is an EU FP7 project that aims to develop combinatorial explosion as the collection of possible input a “computationally feasible, formal model of conceptual spaces is explored, and possible blends are computed. Copyright c 2015, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. 1http://www.coinvent-project.eu/

15 VGDL: The Video Game Description Language allows Weaken the blend by removing/adjusting elements to the specification of simple arcade-style games (Schaul achieve viability. 2014). It was recently developed to support research into ar- Run the blend by elaborating the design to exploit any new tificial intelligence and games, in particular general game- structure which it has introduced. playing agents, and now forms the basis of the General Video Game Playing AI (GVG-AI) Competition (Perez et The following account is a thought experiment, but is, we al. 2015). argue, computationally plausible. We expect that the ma- VGDL conceptualises a game as having five components: jority of games developed by such a process would not be playable, and indeed there will be many partially devel- Sprite Set The entities in the game, their types and proper- oped blends that will fail. For non-trivial examples, each ties. The predefined types are Avatar, FromAvatar, NPC, stage presents a huge number of choices — a combinato- Static, Movable, Resource and Portal, plus a large number rial explosion which any automated designer needs to tackle of specialised subtypes, e.g. RandomNPC. The sprites are in order to extract playable needles out of a haystack of arranged in a hierarchy, where subsprites (here denoted ‘frankengames’. We draw attention below to some of the de- v) inherit types/properties of their ancestors. cisions a computational designer would have to confront. Interaction Set The mechanics of the game. Each interac- The term “designer” is used to refer to this imagined tion rule defines the consequences of a particular pair of agent, but one might also conceive a co-creative approach sprites overlapping. where a human designer takes some of the key decisions, and perhaps takes inspiration from the blends the system Termination Set The win/lose conditions, based on either produces. a timeout or the value of counter, e.g. player health. Previous work in automated game design has suggested Level Mapping A translation from ASCII characters to that there are few intrinsic criteria for game quality (Nel- sprites. son et al. 2015). Playtesting, whether by humans or bots, is an essential component of game design. An implementa- Level Designs Multiple levels can be defined in ASCII, in- tion could use feedback from automated playtesting of the terpreted using the Level Mapping. games to guide this process, as in e.g. (Nielsen et al. 2015a; A VGDL specification defines the first four elements, with 2015b). level designs being specified in a set of accompanying text We completely sidestep the difficult issues of blending files. level designs and blending audiovisual elements, focusing The sprites and rules are described in terms of a prede- instead on blending game entities and rules. Of course, the fined ontology, e.g. a sprite can be a RandomNPC which design of all these elements is intimitely connected, and a wanders around the screen. The default ontology is fairly full account of game blending would encompass these as limited, and the VGDL versions of well-known titles are of- well. ten quite simplistic. Nevertheless, it can be used to define an impressive range of games which present a serious challenge Selection for a general game-playing agent, as demonstrated by the The first question is which games should we blend? Input GVG Competition examples. In the context of game design, game selection can be thought of as taking part in some VGDL presents a large space of possible games. Random wider game design context: perhaps an agent is elaborating instances from within that space are unlikely to be coherant, a particular design, and decides to blend in another game playable games, let alone high quality ones (Barros and To- concept in order to fix a perceived problem with the orig- gelius 2014). Hence it also presents a research challenge for inal. Alternatively, an agent could be simply exploring the automated game design, and researchers are now beginning creative possibilities of mashing together diverse game de- to look at VGDL generation (Nielsen et al. 2015b). signs. For this paper, we have just hand-selected two games Blending VGDL from the GVG Competition example set: Frogger (AKA Taking VGDL specifications as our ‘conceptual spaces’ for Frogs) and Zelda (Perez et al. 2015). Both games are in- games, we wish to show how a computational agent could spired by well-known commercial titles and have simple conceivably create new designs by blending existing game specifications — the VGDL is shown in Figure 1 and 2 re- elements. Our proposed approach closely follows the gen- spectively. The level and visual design are both outside the eral COINVENT process model: scope of this paper, so for simplicity we have omitted the LevelMapping information and img and color prop- Select two input games to blend. erties. We have also renamed a few sprites and made minor Generalise the games by matching selected elements from alterations for clarity, without affecting the game mechanics. each, resulting in a generic game concept. It is possible to intuit a lot about how these games work from the VGDL descriptions: the player controls an Avatar, Blend the games by combining their constituent elements, Missiles travel in given direction etc. Due to lack of space, reformulated in terms of the generic game concept. we do not include a full description of the games: see Schaul Evaluate whether the (perhaps nonsensical) blend is a vi- (2014) for more details, or play them using the Java emulator able game. at https://github.com/EssexUniversityMCTS/gvgai.

16 SpriteSet SpriteSet forest > SpawnPoint stype=log door > Door dense > prob=0.4 cooldown=10 key > Immovable sparse > prob=0.1 cooldown=5 wall > Immovable structure > Immovable sword > Flicker limit=5 singleton=True home > Door movable > water link > ShootAvatar stype=sword wall nokey log > Missile orientation=LEFT speed=0.1 withkey safety > Resource limit=2 monster > RandomNPC frog > MovingAvatar monsterQuick > cooldown=2 truck > Missile img=truck monsterNormal > cooldown=4 rightTruck > orientation=RIGHT monsterSlow > cooldown=8 fastRtruck > speed=0.2 slowRtruck > speed=0.1 InteractionSet leftTruck > orientation=LEFT movable wall > stepBack fastLtruck > speed=0.2 nokey door > stepBack slowLtruck > speed=0.1 door withkey > killSprite scoreChange=1 monster sword > killSprite scoreChange=2 InteractionSet link monster > killSprite scoreChange=-1 home frog > killSprite scoreChange=1 key link > killSprite scoreChange=1 frog log > changeResource resource=safety value=2 nokey key > transformTo stype=withkey frog log > pullWithIt frog wall > stepBack TerminationSet frog water > killIfHasLess resource=safety limit=0 SpriteCounter stype=door win=True frog water > changeResource resource=safety value=-1 SpriteCounter stype=link win=False frog truck > killSprite scoreChange=-2 log EOS > killSprite truck EOS > wrapAround Figure 2: Zelda VGDL speciﬁcation.

TerminationSet SpriteCounter stype=home win=True Generic Frogger Zelda SpriteCounter stype=frog win=False avatar frog link goal home door wall wall wall Figure 1: Frogger VGDL speciﬁcation. enemy truck monster fastREnemy fastRTruck monsterQuick fastLEnemy fastLTruck monsterQuick Generalisation slowREnemy slowRTruck monsterSlow slowLEnemy slowLTruck monsterSlow The generalisation step proposes a generic game concept which can be mapped onto elements of the two input games. Note that both input games have additional unmapped In the case of VGDL, we need to propose one or more sprites, e.g. Zelda’s key and sword don’t correspond to generic sprites, which can then be mapped on to a subset anything in the generic game. of the two input sprite sets. For example, if game X features aliens and game Y ghosts, we could propose a generic Blending the Sprites G enemies signa- game concept with . We then have two An initial version of our blended game — let’s call it Frolda ture morphisms G , i.e. mappings from the sprites of to the — can now be generated. The blended sprite set combines φ(G, X): enemy 7→ alien sprites of the input games: and those from both games, merging sprites identiﬁed by the sig- φ(G, Y ): enemy 7→ ghost . nature morphisms: this is shown in Figure 3. Even for quite small sprite sets the number of possi- In generating this blend, we have retained as much struc- ble mappings is considerable. We can imagine quite radical ture as possible from both games: goalvstructure is mappings, e.g. where a sprite is mapped to both the player based on the door from Zelda, and the monster and avatar and a ghost, or both an alien and an ammo pack, and truck sprite hierarchies have been integrated, retaining the these clearly have a lot of creative potential. However, for rightTruck and leftTruck sprites from Frogger (re- now we make the additional simplifying assumption that named rightEnemy and leftEnemy). the mappings preserve the VGDL sprite categories: Avatar, The input games used two separate mechanisms to con- NPC, Resource etc. So the generic player avatar is mapped trol NPC speed — cooldown (rate of action) and speed to the player avatars in both input games, and so on. (distance traveled per movement action) — hence all ene- In our running example, the designer could propose the mies are slower in the blend than either of the inputs. This following category-preserving generic sprite set that maps is an example of a undesirable interaction between game el- as follows: ements that will need to be dealt with later in the blending

17 forest > SpawnPoint stype=log goal avatar > killSprite scoreChange=1 dense > prob=0.4 cooldown=10 goal withkey > killSprite scoreChange=1 sparse > prob=0.1 cooldown=5 nokey goal > stepBack structure > Immovable key avatar > killSprite scoreChange=1 goal > Door nokey key > transformTo stype=withkey water wall movable wall > stepBack key avatar wall > stepBack sword > Flicker limit=5 singleton=True avatar log > changeResource resource=safety value=2 log > Missile orientation=LEFT speed=0.1 avatar log > pullWithIt safety > Resource limit=2 avatar water > killIfHasLess resource=safety limit=0 movable > avatar water > changeResource resource=safety value=-1 avatar > ShootAvatar stype=sword log EOS > killSprite nokey enemy EOS > wrapAround withkey enemy > Missile enemy sword > killSprite scoreChange=2 rightEnemy > orientation=RIGHT avatar enemy > killSprite scoreChange=-1 fastREnemy > cooldown=2 speed=0.2 monsterNormal sword > killSprite scoreChange=2 slowREnemy > cooldown=8 speed=0.1 avatar monsterNormal > killSprite scoreChange=-1 leftEnemy > orientation=LEFT fastLEnemy > cooldown=2 speed=0.2 slowLEnemy > cooldown=8 speed=0.1 Figure 4: Initial interaction set for Frolda monsterNormal > RandomNPC cooldown=4 whether we need to further weaken the blend to improve the Figure 3: Initial sprite set for Frolda. game’s coherance and playability. First, we can see that the initial interaction set (see Figure 4) contains some redundancies: process. avatar movable Two significant decisions have been made during the • and have the same interaction wall avatar movable sprite set blend: with , and v . Also note that truckvmovable and if trucks are blocked by walls 1. The player is a ShootAvatar, following Zelda. then they will be prevented from wrapping around, i.e. 2. An enemy is a Missile, following Frogger. all the trucks will end up stacked against the wall. Hence, we remove the movable interaction and retain the more This is a form of weakening (see below) which it makes specific rule. sense to apply at this stage, adjusting the blend to ensure avatar withkey the initial blend is valid VGDL. • and have the same interaction with goal withkeyvavatar Some elements of the input games are unaffected by , and . If we retain the more the blend: the walls, the river system from Frogger general rule then the player could always get to the goal, (forest/water/log/safety), and the sword, key making the key redundant. Hence we again choose to re- avatar and monsterNormal from Zelda. tain the more specific rule, and drop the interaction. Blending the Rules Another principle we can apply is parsimony: does the game contain unnecessary elements? This is very much We continue the blend by combining the interaction sets: context-dependent. A designer who was working on Frog- an initial version is shown in Figure 4. All interac- ger, but had blended it with Zelda in order to add a key- door tions can be included without modification, except for mechanism might regard the wandering monsterNormal avatar-enemy, which reduces the score by 2 in Frogger, as an unnecessary addition. Likewise, a Zelda designer who but only 1 in Zelda. By choosing the latter value we again had blended in the ‘crossing the road’ structure from Frog- weaken the blend in order to obtain valid VGDL. ger could regard the log/river system as unwanted. If we are Finally, we blend the termination sets. In this example, interested in mashing up the two games, we could regard the structure happens to be identical: the player loses if the any additional structure outside the core of the blend (i.e. avatar is killed, and wins if the goal is ‘killed’ by being the generic shared structure) as unnecessary. Here we adopt touched by the avatar. the latter position, and remove both the wandering monsters (monsterNormal) and river system. Evaluation/Weakening Finally, we may wish to retune values in the blended The initial blend includes all the sprites, interactions and ter- game, such as the speed of the enemies, which may not be mination conditions from both games, with a unique blended appropriate in their new context. As noted above, the en- structure defined by the chosen signature morphisms. Ad- emies speed in the initial blend is being modulated by both justments were made to the blend to ensure we had valid speed and cooldown values. Here, we decide that we can VGDL, which introduced choices about how the unique use the default cooldown value. blend should be weakened. The next step is to consider A final version of Frolda is shown in Figure 5. The game

18 SpriteSet • Blending home with doorvstructure was structure > Immovable goal > Door straightforward. But what if Frogger had specified wall homevbuilding? One option would be to blend key building and structure during the generalisation sword > Flicker limit=5 singleton=True phase. Another would be to create a new relationship in movable > the blend, e.g. buildingvstructure. A third option avatar > ShootAvatar stype=sword would be to remove, say, building entirely. nokey withkey • When blending the sprite set we could have selected enemy > Missile MoveAvatar instead of ShootAvatar, leaving the rightEnemy > orientation=RIGHT sword without a role in the game, to be later removed dur- fastREnemy > speed=0.2 ing weakening. slowREnemy > speed=0.1 • Choosing RandomNPC enemies instead of Missile leftEnemy > orientation=LEFT would have given us a Zelda-style free-roaming monsters fastLEnemy > speed=0.2 slowLEnemy > speed=0.1 instead of the horizontally-moving trucks that can be used in level designs to create a dangerous ‘road’. InteractionSet goal withkey > killSprite scoreChange=1 Next Steps nokey goal > stepBack The account above has glossed over many details, although key avatar > killSprite scoreChange=1 we hope that each step is computationally plausible. To- nokey key > transformTo stype=withkey avatar wall > stepBack wards this end we intend to implement a basic VGDL avatar enemy > killSprite scoreChange=-1 blender using the Java VGDL library (Perez et al. 2015). enemy EOS > wrapAround We envision the primary research topic for blended games enemy sword > killSprite scoreChange=2 to surround experimentation with heuristics: can we man- age the massive combinatorial possibilities to produce a di- TerminationSet verse range of non-trivial, coherent, novel, playable games? SpriteCounter stype=goal win=True A subsequent phase of work would integrate the imple- SpriteCounter stype=avatar win=False mentation with the COINVENT software stack, as that ma- tures. The COINVENT system currently accepts input in Figure 5: Final specification for Frolda. the Web Ontology Language (OWL) or the Common Al- gebraic Specification Language (CASL), so one likely approach would be to write a translator, for instance between features enemies that move laterally left and right across the VGDL descriptions and OWL. screen, but which can be killed by using a sword, and re- A more immediate next step to support system-building is quires the player to collect the key before reaching the goal. to work out further compelling examples in detail, e.g. Figure 6 shows the game running in the GVG-AI Java em- Pac-Man + Zombies The Survive Zombies game that ships ulator, with our choice of sprite images and level design. with the GVG-AI collection has some similarity to Pac- This level takes advantage of the key-door mechanic to force Man (Perez et al. 2015). Both games give the protago- Frolda to cross the ‘road’ of monsters twice before exiting nist the task of collecting a certain resource while avoid- the level. ing a population of Chasers. However, in Survive Zom- ‘Running the blend’ refers to further elaborating the bies another sprite is introduced (bees) which transforms blended game to take advantage of the synthesis of game the chasers into honey. One possible blend with Pac-Man elements. For example, we may decide to make the key a would allow the protagonist to place fruit that is toxic to Missile so that as well as ‘crossing the road’ of enemies, the ghosts in strategic locations. player has to pick the key out of the moving traffic. Sokoban + Snake The protagonist’s tail grows when boxes are pushed to the goal2. Alternative Blends Sokoban + Sokoban Two copies of Sokoban hooked to- One can imagine many alternative choices to the ones made gether, so that boxes cleared from one side appear on above. Consider these examples, out of the many possible the other. Perhaps best conceived of as a two-player blends: game, with each player in the role of “Maxwell’s demon”. • A generic game which doesn’t distinguish the speed of VGDL would have to be extended to multi-player games. enemies, resulting in single-speed monsters which travel Two other directions that we did not address here in- left and right. clude level design and audiovisual design for games. Rather than directly blending the example levels that accompany • A generic game without a unified class of enemies. Trucks and monsters would both be included in the blend as a 2It seems challenging to implement Snake in VGDL, since the separate sprites, and the designer could run the blend by snake’s tail would have to be made of a chain of chasers that are inventing new interactions for them. programmed to follow each other.

19 Figure 6: A screenshot of Frolda running in the GVG-AI Java emulator. The player (bottom left) has to cross the ‘road’ of monsters to get the key (top) and then cross again before exiting the level (bottom).

VGDL, we propose to express the level idiom in the form ements to achieve these kind of effects. of a declarative level generator, e.g. in ASP (Smith and Brandt and Brandt offer a critique of conceptual blending Mateas 2011), and blend these generators in concert with that can help in thinking about what makes a blended game the rest of game description. Other aspects of visual design playable (Brandt and Brandt 2005). They argue that a purely (e.g. sprites) could be developed through blending, building conceptual approach to blending is not as cognitively useful on recent work in icon blending (Confalonieri et al. 2015). as an approach that considers blends in their communica- Given its focus on game-playing agents, VGDL does not tive context. In the case of games, there is both a narrative currently support any audio that could be blended, although and an interactive component to the communication. This is it would be relatively easy to extend with simple audio ef- consistent with the view from PCG that intrinsic criteria are fects to support research in that direction. insufficient for automated game design, and that some form of playtesting is required (Nelson et al. 2015). Discussion One aspect in which our account is significantly lacking is Conclusions how a computational designer can “run the blend” to take advantage of the new creative opportunities afforded by the We have shown that it is — in principle — possible to pro- newly combined elements. Consider the variant mentioned vide a computational approach to blending games specified above where trucks and monsters remain separate sprites. in VGDL, although many such blends are possible. We high- How should they interact? We expect our initial implemen- lighted the opportunities for emergent meaning making that tation will use quite basic mechanisms, e.g. randomly se- are afforded to game designers using this approach. Never- lecting an event (or no event) for the interaction of the two theless, it is difficult to point to anything resembling elabo- sprites. A more radical move would be to elaborate new me- ration, evaluation, or meaning-making within VGDL games chanics. For instance, monsters could commandeer and drive themselves: for instance, the observation that a given game trucks. To give Frolda a chance to survive in this new Mad is unduly difficult for the player would have to come from Max universe, it may be important to give her the oppor- outside. Automated playtesting could provide some of the tunity to collect a bazooka that can launch missiles at the necessary feedback to guide the blending process, and will trucks. How this kind of elborative process can be realised be integral to the implementation of the approach we have computationally could be a direction for future research. outlined here. The player’s imagination does a significant amount of work to make sense of the blend — perhaps especially in a game with limited graphics. Knowledge of the original Acknowledgments games may help the player envision the new blended sce- nario. Thus, if we recognise the two source games, Frolda Many thanks to Christian Guckelsberger and the anonymous may be envisioned as an anthropomorphic frog equipped reviewers for their comments on an earlier draft of this paper. with a sword. We have only considered structural blending, Joseph Corneli was supported by the European Commision but a broader view could include blending of audiovisual el- via the COINVENT project (FP7 FET-Open grant 611553).

20 References Schorlemmer, M.; Smaill, A.; Kuhnberger,¨ K.; Kutz, O.; Colton, S.; Cambouropoulos, E.; and Pease, A. 2014. COIN- Barros, G., and Togelius, J. 2014. Exploring a large space of VENT: Towards a computational concept invention theory. small games. In Proc. 2014 IEEE Conf. on Computational In Proc. 5th Int. Conf. on Computational Creativity, ICCC Intelligence and Games, CIG 2014. 2014. ACC. http://computationalcreativity.net/. Bou, F.; Corneli, J.; Gomez-Ram´ ´ırez, D.; Maclean, E.; Smith, A. M., and Mateas, M. 2011. Answer set program- Pease, A.; Schorlemmer, M.; and Smaill, A. 2015. The ming for procedural content generation: A design space ap- role of blending in mathematical invention. In Colton, S.; proach. IEEE Trans. Comput. Intellig. and AI in Games Toivonen, H.; Cook, M.; and Ventura, D., eds., Proc. 6th 3(3):187–200. Int. Conf. on Computational Creativity, ICCC 2015. ACC. http://computationalcreativity.net/. Treanor, M.; Blackford, B.; Mateas, M.; and Bogost, I. 2012. Game-o-matic: Generating videogames that represent ideas. Brandt, L., and Brandt, P. A. 2005. Making sense of a blend: In Proc. 3rd Workshop on Procedural Content Generation A cognitive-semiotic approach to metaphor. Annual Review in Games. ACM. of Cognitive Linguistics 3(1):216–249. Cambouropoulos, E.; Kaliakatsos-Papakostas, M.; and Tsougras, C. 2014. An idiom-independent representation of chords for computational music analysis and generation. In Proceeding of the joint 11th Sound and Music Computing Conference (SMC) and 40th International Computer Music Conference (ICMC), Athens, Greece. Confalonieri, R.; Corneli, J.; Pease, A.; Plaza, E.; and Schor- lemmer, M. 2015. Using argumentation to evaluate concept blends in combinatorial creativity. In Colton, S.; Toivo- nen, H.; Cook, M.; and Ventura, D., eds., Proc. 6th Int. Conf. on Computational Creativity, ICCC 2015. ACC. http: //computationalcreativity.net/. Cook, M., and Colton, S. 2014. Ludus ex machina: Build- ing a 3D game designer that competes alongside humans. In Proc. 5th Int. Conf. on Computational Creativity, ICCC 2014. ACC. Goguen, J. 1999. An introduction to algebraic semiotics, with application to user interface design. In Computation for metaphors, analogy, and agents. Springer. 242–291. Nelson, M. J.; Togelius, J.; Browne, C.; and Cook, M. 2015. Rules and mechanics. In Shaker, N.; Togelius, J.; and Nel- son, M. J., eds., Procedural Content Generation in Games. Springer. chapter 6. http://www.pcgbook.com. Nielsen, T. S.; Barros, G.; Togelius, J.; and Nelson, M. J. 2015a. General video game evaluation using relative algo- rithm performance profiles. In Proceedings of EvoApplica- tions 2015. Nielsen, T. S.; Barros, G. A. B.; Togelius, J.; and Nelson, M. J. 2015b. Towards generating arcade game rules with VGDL. In Proc. 2015 IEEE Conf. on Computational Intel- ligence and Games, CIG 2015. IEEE. Ontanõn,´ S., and Plaza, E. 2010. Amalgams: A formal approach for combining multiple case solutions. In IC- CBR10: 18th International Conference on Case-Based Rea- soning, volume 6176 of Lecture Notes in Artificial Intelli- gence. Springer. 257–271. Perez, D.; Samothrakis, S.; Togelius, J.; Schaul, T.; and Lu- cas, S. 2015. The General Video Game AI Competition. http://www.gvgai.net/. Schaul, T. 2014. An extensible description language for video games. IEEE Trans. Comput. Intellig. and AI in Games 6(4):325–331.